MMADPE: drug repositioning based on multi-hop graph Mamba aggregation with dual-modality graph positional encoding.

Abstract

Drug repositioning (DR) has emerged as a critical research strategy for uncovering novel therapeutic applications of existing drugs, demonstrating considerable clinical significance. Despite promising advancements in computational methods for predicting drug-disease associations, most algorithms exhibit three major limitations. First, they inadequately capture high-order relationships within drug-disease networks. Second, they fail to concurrently model both local interaction strengths and global network topologies. Most importantly, current models lack biological interpretability and are incapable of extracting meaningful biological patterns from numerical data. To address these challenges, we propose a novel drug repositioning framework, MMADPE. This framework constructs a similarity network supporting multi-hop aggregation and leverages the linear complexity of Graph Mamba to efficiently integrate multi-order neighborhood information, thereby significantly enhancing the modeling of long-range drug-disease interactions. Subsequently, a dual-modal graph positional encoding is employed, capturing global network topology via Laplacian eigenvectors and characterizing local node association strengths through random walk statistics. Finally, the framework incorporates a GraphGPS hybrid architecture that fuses gated graph convolution with Transformer attention mechanisms to extract molecular biochemical features and their semantic relationships, achieving a deep integration of topological structures and biological semantics. Extensive experiments on three benchmark datasets demonstrate that MMADPE consistently outperforms state-of-the-art methods in drug repositioning tasks. Notably, case studies on two common diseases combined with molecular docking experiments not only validate the effectiveness of our approach but also provide novel mechanistic insights into MMADPE's ability to identify previously unrecognized drug-disease associations.